基于显式运动建模的视频伪装目标检测  

Explicit motion handling for video camouflaged object detection

作　　者：肖涛 章超[2,3] 傅可人 XIAO Tao;ZHANG Chao;FU Keren(School of Computer Science,Sichuan University,Chengdu 610065,China;Sichuan Police College,Luzhou 646000,China;Intelligent Policing Key Laboratory of Sichuan Province,Luzhou 646000,China;National Key Laboratory of Fundamental Science on Synthetic Vision,Sichuan University,Chengdu 610064,China)

机构地区：[1]四川大学计算机学院,成都610065 [2]四川警察学院,泸州646000 [3]智能警务四川省重点实验室,泸州646000 [4]四川大学视觉合成图形图像技术国家级重点实验室,成都610064

出　　处：《上海理工大学学报》2024年第2期120-128,共9页Journal of University of Shanghai For Science and Technology

基　　金：国家自然科学基金资助项目(62176169);智能警务四川省重点实验室资助项目(ZNJW2022KFMS001)。

摘　　要：目前的视频伪装目标检测方法通常采用隐式运动建模或直接输入存在噪声的离线光流图来获取运动线索,这会影响模型性能。为了解决这一问题,提出一种新的基于显式运动建模的视频伪装目标检测框架,称为SMHNet。首先,该框架将显式运动建模与伪装目标检测联合在同一个框架中进行学习。然后利用特征双向更新模块实现两个分支的双向交互更新,相互补充、优化和纠错,输出光流估计结果和目标检测图。此外,为了解决缺少光流真值图这一问题,采用自监督策略对显式运动建模分支进行监督。在两个数据集上的对比实验结果表明,SMHNet有效地提高了视频场景中伪装目标检测的性能。Earlier video camouflaged object detection methods often exploit motion cues by implicit motion modeling or directly inputting offline optical flow maps with noise,which affects model performance.To address the effective utilization of motion cues,an explicit motion modeling framework for video camouflaged object detection,called SMHNet,was proposed.First,an explicit motion modeling branch and a camouflaged object detection branch were jointly learned in the same framework.Then,the two branches were updated bidirectionally using a bidirectional feature updating module.The two branches performed mutual optimization and error correction to output optical flow estimation results and object detection maps.In addition,to address the lack of ground truth optical flow,a self-supervised strategy was adopted to supervise the explicit motion modeling branch.Comparison experiments on two benchmark datasets show that SMHNet effectively improves the performance of video camouflaged object detection.

关 键 词：视频伪装目标检测 显式运动建模 光流 自监督 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]